Projects

ExplaiNN Discovering patterns of neuron activations

We propose an algorithm that allows to find rules of neuron activations that describe how information is encoded and flowswithin a Neural Network, and allows to super-charge prototyping.

Plant'n'Seek Can you find the winning ticket?

We study whether current limitations in network pruning are algorithmic or foundational by crafting hidden lottery tickets in large neural networks for bechnmarking.

Premise Label-descriptive patterns

We propose the algorithm Premise to discover label-descriptive patterns, which we show to be highly useful in characterizing classification errors in multimodal and language models.

Binaps Differentiable pattern mining

We propose a fully differentiable approach based on binary neural networks that allows to discover patterns through a neuro-symbolic architecture.

Diffnaps Finding class-specific pattern through efficient neural search

We propose an efficient approach to learn class-specific pattern through a novel neuro-symbolic architecture, which is fully interpretable yet scalable to hundreds of thousands of features.

Phoenix Modeling gene regulatory dynamics

We propose an easy-to-interpret and highly efficient neural approach to model gene regulatory dynamics that allows to scale to large gene expression data.

DASH Prior-informed pruning of neural networks

We suggest a pruning approach that takes into account prior domain-knowledge to align the pruned network structure with that knowledge, thus improving pruning efficiency and quality of encoded information.

Mercat Faithful low-dimensional embeddings through angle preservation

To overcome consistent problems in existing low-dimensional embedding methodology, we propose to reframe the problem reconstructing angles between any point triplet, instead of reconstructing distances.